xref: /linux/mm/userfaultfd.c (revision 8804d970fab45726b3c7cd7f240b31122aa94219)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  *  mm/userfaultfd.c
4  *
5  *  Copyright (C) 2015  Red Hat, Inc.
6  */
7 
8 #include <linux/mm.h>
9 #include <linux/sched/signal.h>
10 #include <linux/pagemap.h>
11 #include <linux/rmap.h>
12 #include <linux/swap.h>
13 #include <linux/swapops.h>
14 #include <linux/userfaultfd_k.h>
15 #include <linux/mmu_notifier.h>
16 #include <linux/hugetlb.h>
17 #include <linux/shmem_fs.h>
18 #include <asm/tlbflush.h>
19 #include <asm/tlb.h>
20 #include "internal.h"
21 #include "swap.h"
22 
23 static __always_inline
validate_dst_vma(struct vm_area_struct * dst_vma,unsigned long dst_end)24 bool validate_dst_vma(struct vm_area_struct *dst_vma, unsigned long dst_end)
25 {
26 	/* Make sure that the dst range is fully within dst_vma. */
27 	if (dst_end > dst_vma->vm_end)
28 		return false;
29 
30 	/*
31 	 * Check the vma is registered in uffd, this is required to
32 	 * enforce the VM_MAYWRITE check done at uffd registration
33 	 * time.
34 	 */
35 	if (!dst_vma->vm_userfaultfd_ctx.ctx)
36 		return false;
37 
38 	return true;
39 }
40 
41 static __always_inline
find_vma_and_prepare_anon(struct mm_struct * mm,unsigned long addr)42 struct vm_area_struct *find_vma_and_prepare_anon(struct mm_struct *mm,
43 						 unsigned long addr)
44 {
45 	struct vm_area_struct *vma;
46 
47 	mmap_assert_locked(mm);
48 	vma = vma_lookup(mm, addr);
49 	if (!vma)
50 		vma = ERR_PTR(-ENOENT);
51 	else if (!(vma->vm_flags & VM_SHARED) &&
52 		 unlikely(anon_vma_prepare(vma)))
53 		vma = ERR_PTR(-ENOMEM);
54 
55 	return vma;
56 }
57 
58 #ifdef CONFIG_PER_VMA_LOCK
59 /*
60  * uffd_lock_vma() - Lookup and lock vma corresponding to @address.
61  * @mm: mm to search vma in.
62  * @address: address that the vma should contain.
63  *
64  * Should be called without holding mmap_lock.
65  *
66  * Return: A locked vma containing @address, -ENOENT if no vma is found, or
67  * -ENOMEM if anon_vma couldn't be allocated.
68  */
uffd_lock_vma(struct mm_struct * mm,unsigned long address)69 static struct vm_area_struct *uffd_lock_vma(struct mm_struct *mm,
70 				       unsigned long address)
71 {
72 	struct vm_area_struct *vma;
73 
74 	vma = lock_vma_under_rcu(mm, address);
75 	if (vma) {
76 		/*
77 		 * We know we're going to need to use anon_vma, so check
78 		 * that early.
79 		 */
80 		if (!(vma->vm_flags & VM_SHARED) && unlikely(!vma->anon_vma))
81 			vma_end_read(vma);
82 		else
83 			return vma;
84 	}
85 
86 	mmap_read_lock(mm);
87 	vma = find_vma_and_prepare_anon(mm, address);
88 	if (!IS_ERR(vma)) {
89 		bool locked = vma_start_read_locked(vma);
90 
91 		if (!locked)
92 			vma = ERR_PTR(-EAGAIN);
93 	}
94 
95 	mmap_read_unlock(mm);
96 	return vma;
97 }
98 
uffd_mfill_lock(struct mm_struct * dst_mm,unsigned long dst_start,unsigned long len)99 static struct vm_area_struct *uffd_mfill_lock(struct mm_struct *dst_mm,
100 					      unsigned long dst_start,
101 					      unsigned long len)
102 {
103 	struct vm_area_struct *dst_vma;
104 
105 	dst_vma = uffd_lock_vma(dst_mm, dst_start);
106 	if (IS_ERR(dst_vma) || validate_dst_vma(dst_vma, dst_start + len))
107 		return dst_vma;
108 
109 	vma_end_read(dst_vma);
110 	return ERR_PTR(-ENOENT);
111 }
112 
uffd_mfill_unlock(struct vm_area_struct * vma)113 static void uffd_mfill_unlock(struct vm_area_struct *vma)
114 {
115 	vma_end_read(vma);
116 }
117 
118 #else
119 
uffd_mfill_lock(struct mm_struct * dst_mm,unsigned long dst_start,unsigned long len)120 static struct vm_area_struct *uffd_mfill_lock(struct mm_struct *dst_mm,
121 					      unsigned long dst_start,
122 					      unsigned long len)
123 {
124 	struct vm_area_struct *dst_vma;
125 
126 	mmap_read_lock(dst_mm);
127 	dst_vma = find_vma_and_prepare_anon(dst_mm, dst_start);
128 	if (IS_ERR(dst_vma))
129 		goto out_unlock;
130 
131 	if (validate_dst_vma(dst_vma, dst_start + len))
132 		return dst_vma;
133 
134 	dst_vma = ERR_PTR(-ENOENT);
135 out_unlock:
136 	mmap_read_unlock(dst_mm);
137 	return dst_vma;
138 }
139 
uffd_mfill_unlock(struct vm_area_struct * vma)140 static void uffd_mfill_unlock(struct vm_area_struct *vma)
141 {
142 	mmap_read_unlock(vma->vm_mm);
143 }
144 #endif
145 
146 /* Check if dst_addr is outside of file's size. Must be called with ptl held. */
mfill_file_over_size(struct vm_area_struct * dst_vma,unsigned long dst_addr)147 static bool mfill_file_over_size(struct vm_area_struct *dst_vma,
148 				 unsigned long dst_addr)
149 {
150 	struct inode *inode;
151 	pgoff_t offset, max_off;
152 
153 	if (!dst_vma->vm_file)
154 		return false;
155 
156 	inode = dst_vma->vm_file->f_inode;
157 	offset = linear_page_index(dst_vma, dst_addr);
158 	max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
159 	return offset >= max_off;
160 }
161 
162 /*
163  * Install PTEs, to map dst_addr (within dst_vma) to page.
164  *
165  * This function handles both MCOPY_ATOMIC_NORMAL and _CONTINUE for both shmem
166  * and anon, and for both shared and private VMAs.
167  */
mfill_atomic_install_pte(pmd_t * dst_pmd,struct vm_area_struct * dst_vma,unsigned long dst_addr,struct page * page,bool newly_allocated,uffd_flags_t flags)168 int mfill_atomic_install_pte(pmd_t *dst_pmd,
169 			     struct vm_area_struct *dst_vma,
170 			     unsigned long dst_addr, struct page *page,
171 			     bool newly_allocated, uffd_flags_t flags)
172 {
173 	int ret;
174 	struct mm_struct *dst_mm = dst_vma->vm_mm;
175 	pte_t _dst_pte, *dst_pte;
176 	bool writable = dst_vma->vm_flags & VM_WRITE;
177 	bool vm_shared = dst_vma->vm_flags & VM_SHARED;
178 	spinlock_t *ptl;
179 	struct folio *folio = page_folio(page);
180 	bool page_in_cache = folio_mapping(folio);
181 
182 	_dst_pte = mk_pte(page, dst_vma->vm_page_prot);
183 	_dst_pte = pte_mkdirty(_dst_pte);
184 	if (page_in_cache && !vm_shared)
185 		writable = false;
186 	if (writable)
187 		_dst_pte = pte_mkwrite(_dst_pte, dst_vma);
188 	if (flags & MFILL_ATOMIC_WP)
189 		_dst_pte = pte_mkuffd_wp(_dst_pte);
190 
191 	ret = -EAGAIN;
192 	dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
193 	if (!dst_pte)
194 		goto out;
195 
196 	if (mfill_file_over_size(dst_vma, dst_addr)) {
197 		ret = -EFAULT;
198 		goto out_unlock;
199 	}
200 
201 	ret = -EEXIST;
202 	/*
203 	 * We allow to overwrite a pte marker: consider when both MISSING|WP
204 	 * registered, we firstly wr-protect a none pte which has no page cache
205 	 * page backing it, then access the page.
206 	 */
207 	if (!pte_none_mostly(ptep_get(dst_pte)))
208 		goto out_unlock;
209 
210 	if (page_in_cache) {
211 		/* Usually, cache pages are already added to LRU */
212 		if (newly_allocated)
213 			folio_add_lru(folio);
214 		folio_add_file_rmap_pte(folio, page, dst_vma);
215 	} else {
216 		folio_add_new_anon_rmap(folio, dst_vma, dst_addr, RMAP_EXCLUSIVE);
217 		folio_add_lru_vma(folio, dst_vma);
218 	}
219 
220 	/*
221 	 * Must happen after rmap, as mm_counter() checks mapping (via
222 	 * PageAnon()), which is set by __page_set_anon_rmap().
223 	 */
224 	inc_mm_counter(dst_mm, mm_counter(folio));
225 
226 	set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
227 
228 	/* No need to invalidate - it was non-present before */
229 	update_mmu_cache(dst_vma, dst_addr, dst_pte);
230 	ret = 0;
231 out_unlock:
232 	pte_unmap_unlock(dst_pte, ptl);
233 out:
234 	return ret;
235 }
236 
mfill_atomic_pte_copy(pmd_t * dst_pmd,struct vm_area_struct * dst_vma,unsigned long dst_addr,unsigned long src_addr,uffd_flags_t flags,struct folio ** foliop)237 static int mfill_atomic_pte_copy(pmd_t *dst_pmd,
238 				 struct vm_area_struct *dst_vma,
239 				 unsigned long dst_addr,
240 				 unsigned long src_addr,
241 				 uffd_flags_t flags,
242 				 struct folio **foliop)
243 {
244 	void *kaddr;
245 	int ret;
246 	struct folio *folio;
247 
248 	if (!*foliop) {
249 		ret = -ENOMEM;
250 		folio = vma_alloc_folio(GFP_HIGHUSER_MOVABLE, 0, dst_vma,
251 					dst_addr);
252 		if (!folio)
253 			goto out;
254 
255 		kaddr = kmap_local_folio(folio, 0);
256 		/*
257 		 * The read mmap_lock is held here.  Despite the
258 		 * mmap_lock being read recursive a deadlock is still
259 		 * possible if a writer has taken a lock.  For example:
260 		 *
261 		 * process A thread 1 takes read lock on own mmap_lock
262 		 * process A thread 2 calls mmap, blocks taking write lock
263 		 * process B thread 1 takes page fault, read lock on own mmap lock
264 		 * process B thread 2 calls mmap, blocks taking write lock
265 		 * process A thread 1 blocks taking read lock on process B
266 		 * process B thread 1 blocks taking read lock on process A
267 		 *
268 		 * Disable page faults to prevent potential deadlock
269 		 * and retry the copy outside the mmap_lock.
270 		 */
271 		pagefault_disable();
272 		ret = copy_from_user(kaddr, (const void __user *) src_addr,
273 				     PAGE_SIZE);
274 		pagefault_enable();
275 		kunmap_local(kaddr);
276 
277 		/* fallback to copy_from_user outside mmap_lock */
278 		if (unlikely(ret)) {
279 			ret = -ENOENT;
280 			*foliop = folio;
281 			/* don't free the page */
282 			goto out;
283 		}
284 
285 		flush_dcache_folio(folio);
286 	} else {
287 		folio = *foliop;
288 		*foliop = NULL;
289 	}
290 
291 	/*
292 	 * The memory barrier inside __folio_mark_uptodate makes sure that
293 	 * preceding stores to the page contents become visible before
294 	 * the set_pte_at() write.
295 	 */
296 	__folio_mark_uptodate(folio);
297 
298 	ret = -ENOMEM;
299 	if (mem_cgroup_charge(folio, dst_vma->vm_mm, GFP_KERNEL))
300 		goto out_release;
301 
302 	ret = mfill_atomic_install_pte(dst_pmd, dst_vma, dst_addr,
303 				       &folio->page, true, flags);
304 	if (ret)
305 		goto out_release;
306 out:
307 	return ret;
308 out_release:
309 	folio_put(folio);
310 	goto out;
311 }
312 
mfill_atomic_pte_zeroed_folio(pmd_t * dst_pmd,struct vm_area_struct * dst_vma,unsigned long dst_addr)313 static int mfill_atomic_pte_zeroed_folio(pmd_t *dst_pmd,
314 					 struct vm_area_struct *dst_vma,
315 					 unsigned long dst_addr)
316 {
317 	struct folio *folio;
318 	int ret = -ENOMEM;
319 
320 	folio = vma_alloc_zeroed_movable_folio(dst_vma, dst_addr);
321 	if (!folio)
322 		return ret;
323 
324 	if (mem_cgroup_charge(folio, dst_vma->vm_mm, GFP_KERNEL))
325 		goto out_put;
326 
327 	/*
328 	 * The memory barrier inside __folio_mark_uptodate makes sure that
329 	 * zeroing out the folio become visible before mapping the page
330 	 * using set_pte_at(). See do_anonymous_page().
331 	 */
332 	__folio_mark_uptodate(folio);
333 
334 	ret = mfill_atomic_install_pte(dst_pmd, dst_vma, dst_addr,
335 				       &folio->page, true, 0);
336 	if (ret)
337 		goto out_put;
338 
339 	return 0;
340 out_put:
341 	folio_put(folio);
342 	return ret;
343 }
344 
mfill_atomic_pte_zeropage(pmd_t * dst_pmd,struct vm_area_struct * dst_vma,unsigned long dst_addr)345 static int mfill_atomic_pte_zeropage(pmd_t *dst_pmd,
346 				     struct vm_area_struct *dst_vma,
347 				     unsigned long dst_addr)
348 {
349 	pte_t _dst_pte, *dst_pte;
350 	spinlock_t *ptl;
351 	int ret;
352 
353 	if (mm_forbids_zeropage(dst_vma->vm_mm))
354 		return mfill_atomic_pte_zeroed_folio(dst_pmd, dst_vma, dst_addr);
355 
356 	_dst_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
357 					 dst_vma->vm_page_prot));
358 	ret = -EAGAIN;
359 	dst_pte = pte_offset_map_lock(dst_vma->vm_mm, dst_pmd, dst_addr, &ptl);
360 	if (!dst_pte)
361 		goto out;
362 	if (mfill_file_over_size(dst_vma, dst_addr)) {
363 		ret = -EFAULT;
364 		goto out_unlock;
365 	}
366 	ret = -EEXIST;
367 	if (!pte_none(ptep_get(dst_pte)))
368 		goto out_unlock;
369 	set_pte_at(dst_vma->vm_mm, dst_addr, dst_pte, _dst_pte);
370 	/* No need to invalidate - it was non-present before */
371 	update_mmu_cache(dst_vma, dst_addr, dst_pte);
372 	ret = 0;
373 out_unlock:
374 	pte_unmap_unlock(dst_pte, ptl);
375 out:
376 	return ret;
377 }
378 
379 /* Handles UFFDIO_CONTINUE for all shmem VMAs (shared or private). */
mfill_atomic_pte_continue(pmd_t * dst_pmd,struct vm_area_struct * dst_vma,unsigned long dst_addr,uffd_flags_t flags)380 static int mfill_atomic_pte_continue(pmd_t *dst_pmd,
381 				     struct vm_area_struct *dst_vma,
382 				     unsigned long dst_addr,
383 				     uffd_flags_t flags)
384 {
385 	struct inode *inode = file_inode(dst_vma->vm_file);
386 	pgoff_t pgoff = linear_page_index(dst_vma, dst_addr);
387 	struct folio *folio;
388 	struct page *page;
389 	int ret;
390 
391 	ret = shmem_get_folio(inode, pgoff, 0, &folio, SGP_NOALLOC);
392 	/* Our caller expects us to return -EFAULT if we failed to find folio */
393 	if (ret == -ENOENT)
394 		ret = -EFAULT;
395 	if (ret)
396 		goto out;
397 	if (!folio) {
398 		ret = -EFAULT;
399 		goto out;
400 	}
401 
402 	page = folio_file_page(folio, pgoff);
403 	if (PageHWPoison(page)) {
404 		ret = -EIO;
405 		goto out_release;
406 	}
407 
408 	ret = mfill_atomic_install_pte(dst_pmd, dst_vma, dst_addr,
409 				       page, false, flags);
410 	if (ret)
411 		goto out_release;
412 
413 	folio_unlock(folio);
414 	ret = 0;
415 out:
416 	return ret;
417 out_release:
418 	folio_unlock(folio);
419 	folio_put(folio);
420 	goto out;
421 }
422 
423 /* Handles UFFDIO_POISON for all non-hugetlb VMAs. */
mfill_atomic_pte_poison(pmd_t * dst_pmd,struct vm_area_struct * dst_vma,unsigned long dst_addr,uffd_flags_t flags)424 static int mfill_atomic_pte_poison(pmd_t *dst_pmd,
425 				   struct vm_area_struct *dst_vma,
426 				   unsigned long dst_addr,
427 				   uffd_flags_t flags)
428 {
429 	int ret;
430 	struct mm_struct *dst_mm = dst_vma->vm_mm;
431 	pte_t _dst_pte, *dst_pte;
432 	spinlock_t *ptl;
433 
434 	_dst_pte = make_pte_marker(PTE_MARKER_POISONED);
435 	ret = -EAGAIN;
436 	dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
437 	if (!dst_pte)
438 		goto out;
439 
440 	if (mfill_file_over_size(dst_vma, dst_addr)) {
441 		ret = -EFAULT;
442 		goto out_unlock;
443 	}
444 
445 	ret = -EEXIST;
446 	/* Refuse to overwrite any PTE, even a PTE marker (e.g. UFFD WP). */
447 	if (!pte_none(ptep_get(dst_pte)))
448 		goto out_unlock;
449 
450 	set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
451 
452 	/* No need to invalidate - it was non-present before */
453 	update_mmu_cache(dst_vma, dst_addr, dst_pte);
454 	ret = 0;
455 out_unlock:
456 	pte_unmap_unlock(dst_pte, ptl);
457 out:
458 	return ret;
459 }
460 
mm_alloc_pmd(struct mm_struct * mm,unsigned long address)461 static pmd_t *mm_alloc_pmd(struct mm_struct *mm, unsigned long address)
462 {
463 	pgd_t *pgd;
464 	p4d_t *p4d;
465 	pud_t *pud;
466 
467 	pgd = pgd_offset(mm, address);
468 	p4d = p4d_alloc(mm, pgd, address);
469 	if (!p4d)
470 		return NULL;
471 	pud = pud_alloc(mm, p4d, address);
472 	if (!pud)
473 		return NULL;
474 	/*
475 	 * Note that we didn't run this because the pmd was
476 	 * missing, the *pmd may be already established and in
477 	 * turn it may also be a trans_huge_pmd.
478 	 */
479 	return pmd_alloc(mm, pud, address);
480 }
481 
482 #ifdef CONFIG_HUGETLB_PAGE
483 /*
484  * mfill_atomic processing for HUGETLB vmas.  Note that this routine is
485  * called with either vma-lock or mmap_lock held, it will release the lock
486  * before returning.
487  */
mfill_atomic_hugetlb(struct userfaultfd_ctx * ctx,struct vm_area_struct * dst_vma,unsigned long dst_start,unsigned long src_start,unsigned long len,uffd_flags_t flags)488 static __always_inline ssize_t mfill_atomic_hugetlb(
489 					      struct userfaultfd_ctx *ctx,
490 					      struct vm_area_struct *dst_vma,
491 					      unsigned long dst_start,
492 					      unsigned long src_start,
493 					      unsigned long len,
494 					      uffd_flags_t flags)
495 {
496 	struct mm_struct *dst_mm = dst_vma->vm_mm;
497 	ssize_t err;
498 	pte_t *dst_pte;
499 	unsigned long src_addr, dst_addr;
500 	long copied;
501 	struct folio *folio;
502 	unsigned long vma_hpagesize;
503 	pgoff_t idx;
504 	u32 hash;
505 	struct address_space *mapping;
506 
507 	/*
508 	 * There is no default zero huge page for all huge page sizes as
509 	 * supported by hugetlb.  A PMD_SIZE huge pages may exist as used
510 	 * by THP.  Since we can not reliably insert a zero page, this
511 	 * feature is not supported.
512 	 */
513 	if (uffd_flags_mode_is(flags, MFILL_ATOMIC_ZEROPAGE)) {
514 		up_read(&ctx->map_changing_lock);
515 		uffd_mfill_unlock(dst_vma);
516 		return -EINVAL;
517 	}
518 
519 	src_addr = src_start;
520 	dst_addr = dst_start;
521 	copied = 0;
522 	folio = NULL;
523 	vma_hpagesize = vma_kernel_pagesize(dst_vma);
524 
525 	/*
526 	 * Validate alignment based on huge page size
527 	 */
528 	err = -EINVAL;
529 	if (dst_start & (vma_hpagesize - 1) || len & (vma_hpagesize - 1))
530 		goto out_unlock;
531 
532 retry:
533 	/*
534 	 * On routine entry dst_vma is set.  If we had to drop mmap_lock and
535 	 * retry, dst_vma will be set to NULL and we must lookup again.
536 	 */
537 	if (!dst_vma) {
538 		dst_vma = uffd_mfill_lock(dst_mm, dst_start, len);
539 		if (IS_ERR(dst_vma)) {
540 			err = PTR_ERR(dst_vma);
541 			goto out;
542 		}
543 
544 		err = -ENOENT;
545 		if (!is_vm_hugetlb_page(dst_vma))
546 			goto out_unlock_vma;
547 
548 		err = -EINVAL;
549 		if (vma_hpagesize != vma_kernel_pagesize(dst_vma))
550 			goto out_unlock_vma;
551 
552 		/*
553 		 * If memory mappings are changing because of non-cooperative
554 		 * operation (e.g. mremap) running in parallel, bail out and
555 		 * request the user to retry later
556 		 */
557 		down_read(&ctx->map_changing_lock);
558 		err = -EAGAIN;
559 		if (atomic_read(&ctx->mmap_changing))
560 			goto out_unlock;
561 	}
562 
563 	while (src_addr < src_start + len) {
564 		VM_WARN_ON_ONCE(dst_addr >= dst_start + len);
565 
566 		/*
567 		 * Serialize via vma_lock and hugetlb_fault_mutex.
568 		 * vma_lock ensures the dst_pte remains valid even
569 		 * in the case of shared pmds.  fault mutex prevents
570 		 * races with other faulting threads.
571 		 */
572 		idx = linear_page_index(dst_vma, dst_addr);
573 		mapping = dst_vma->vm_file->f_mapping;
574 		hash = hugetlb_fault_mutex_hash(mapping, idx);
575 		mutex_lock(&hugetlb_fault_mutex_table[hash]);
576 		hugetlb_vma_lock_read(dst_vma);
577 
578 		err = -ENOMEM;
579 		dst_pte = huge_pte_alloc(dst_mm, dst_vma, dst_addr, vma_hpagesize);
580 		if (!dst_pte) {
581 			hugetlb_vma_unlock_read(dst_vma);
582 			mutex_unlock(&hugetlb_fault_mutex_table[hash]);
583 			goto out_unlock;
584 		}
585 
586 		if (!uffd_flags_mode_is(flags, MFILL_ATOMIC_CONTINUE) &&
587 		    !huge_pte_none_mostly(huge_ptep_get(dst_mm, dst_addr, dst_pte))) {
588 			err = -EEXIST;
589 			hugetlb_vma_unlock_read(dst_vma);
590 			mutex_unlock(&hugetlb_fault_mutex_table[hash]);
591 			goto out_unlock;
592 		}
593 
594 		err = hugetlb_mfill_atomic_pte(dst_pte, dst_vma, dst_addr,
595 					       src_addr, flags, &folio);
596 
597 		hugetlb_vma_unlock_read(dst_vma);
598 		mutex_unlock(&hugetlb_fault_mutex_table[hash]);
599 
600 		cond_resched();
601 
602 		if (unlikely(err == -ENOENT)) {
603 			up_read(&ctx->map_changing_lock);
604 			uffd_mfill_unlock(dst_vma);
605 			VM_WARN_ON_ONCE(!folio);
606 
607 			err = copy_folio_from_user(folio,
608 						   (const void __user *)src_addr, true);
609 			if (unlikely(err)) {
610 				err = -EFAULT;
611 				goto out;
612 			}
613 
614 			dst_vma = NULL;
615 			goto retry;
616 		} else
617 			VM_WARN_ON_ONCE(folio);
618 
619 		if (!err) {
620 			dst_addr += vma_hpagesize;
621 			src_addr += vma_hpagesize;
622 			copied += vma_hpagesize;
623 
624 			if (fatal_signal_pending(current))
625 				err = -EINTR;
626 		}
627 		if (err)
628 			break;
629 	}
630 
631 out_unlock:
632 	up_read(&ctx->map_changing_lock);
633 out_unlock_vma:
634 	uffd_mfill_unlock(dst_vma);
635 out:
636 	if (folio)
637 		folio_put(folio);
638 	VM_WARN_ON_ONCE(copied < 0);
639 	VM_WARN_ON_ONCE(err > 0);
640 	VM_WARN_ON_ONCE(!copied && !err);
641 	return copied ? copied : err;
642 }
643 #else /* !CONFIG_HUGETLB_PAGE */
644 /* fail at build time if gcc attempts to use this */
645 extern ssize_t mfill_atomic_hugetlb(struct userfaultfd_ctx *ctx,
646 				    struct vm_area_struct *dst_vma,
647 				    unsigned long dst_start,
648 				    unsigned long src_start,
649 				    unsigned long len,
650 				    uffd_flags_t flags);
651 #endif /* CONFIG_HUGETLB_PAGE */
652 
mfill_atomic_pte(pmd_t * dst_pmd,struct vm_area_struct * dst_vma,unsigned long dst_addr,unsigned long src_addr,uffd_flags_t flags,struct folio ** foliop)653 static __always_inline ssize_t mfill_atomic_pte(pmd_t *dst_pmd,
654 						struct vm_area_struct *dst_vma,
655 						unsigned long dst_addr,
656 						unsigned long src_addr,
657 						uffd_flags_t flags,
658 						struct folio **foliop)
659 {
660 	ssize_t err;
661 
662 	if (uffd_flags_mode_is(flags, MFILL_ATOMIC_CONTINUE)) {
663 		return mfill_atomic_pte_continue(dst_pmd, dst_vma,
664 						 dst_addr, flags);
665 	} else if (uffd_flags_mode_is(flags, MFILL_ATOMIC_POISON)) {
666 		return mfill_atomic_pte_poison(dst_pmd, dst_vma,
667 					       dst_addr, flags);
668 	}
669 
670 	/*
671 	 * The normal page fault path for a shmem will invoke the
672 	 * fault, fill the hole in the file and COW it right away. The
673 	 * result generates plain anonymous memory. So when we are
674 	 * asked to fill an hole in a MAP_PRIVATE shmem mapping, we'll
675 	 * generate anonymous memory directly without actually filling
676 	 * the hole. For the MAP_PRIVATE case the robustness check
677 	 * only happens in the pagetable (to verify it's still none)
678 	 * and not in the radix tree.
679 	 */
680 	if (!(dst_vma->vm_flags & VM_SHARED)) {
681 		if (uffd_flags_mode_is(flags, MFILL_ATOMIC_COPY))
682 			err = mfill_atomic_pte_copy(dst_pmd, dst_vma,
683 						    dst_addr, src_addr,
684 						    flags, foliop);
685 		else
686 			err = mfill_atomic_pte_zeropage(dst_pmd,
687 						 dst_vma, dst_addr);
688 	} else {
689 		err = shmem_mfill_atomic_pte(dst_pmd, dst_vma,
690 					     dst_addr, src_addr,
691 					     flags, foliop);
692 	}
693 
694 	return err;
695 }
696 
mfill_atomic(struct userfaultfd_ctx * ctx,unsigned long dst_start,unsigned long src_start,unsigned long len,uffd_flags_t flags)697 static __always_inline ssize_t mfill_atomic(struct userfaultfd_ctx *ctx,
698 					    unsigned long dst_start,
699 					    unsigned long src_start,
700 					    unsigned long len,
701 					    uffd_flags_t flags)
702 {
703 	struct mm_struct *dst_mm = ctx->mm;
704 	struct vm_area_struct *dst_vma;
705 	ssize_t err;
706 	pmd_t *dst_pmd;
707 	unsigned long src_addr, dst_addr;
708 	long copied;
709 	struct folio *folio;
710 
711 	/*
712 	 * Sanitize the command parameters:
713 	 */
714 	VM_WARN_ON_ONCE(dst_start & ~PAGE_MASK);
715 	VM_WARN_ON_ONCE(len & ~PAGE_MASK);
716 
717 	/* Does the address range wrap, or is the span zero-sized? */
718 	VM_WARN_ON_ONCE(src_start + len <= src_start);
719 	VM_WARN_ON_ONCE(dst_start + len <= dst_start);
720 
721 	src_addr = src_start;
722 	dst_addr = dst_start;
723 	copied = 0;
724 	folio = NULL;
725 retry:
726 	/*
727 	 * Make sure the vma is not shared, that the dst range is
728 	 * both valid and fully within a single existing vma.
729 	 */
730 	dst_vma = uffd_mfill_lock(dst_mm, dst_start, len);
731 	if (IS_ERR(dst_vma)) {
732 		err = PTR_ERR(dst_vma);
733 		goto out;
734 	}
735 
736 	/*
737 	 * If memory mappings are changing because of non-cooperative
738 	 * operation (e.g. mremap) running in parallel, bail out and
739 	 * request the user to retry later
740 	 */
741 	down_read(&ctx->map_changing_lock);
742 	err = -EAGAIN;
743 	if (atomic_read(&ctx->mmap_changing))
744 		goto out_unlock;
745 
746 	err = -EINVAL;
747 	/*
748 	 * shmem_zero_setup is invoked in mmap for MAP_ANONYMOUS|MAP_SHARED but
749 	 * it will overwrite vm_ops, so vma_is_anonymous must return false.
750 	 */
751 	if (WARN_ON_ONCE(vma_is_anonymous(dst_vma) &&
752 	    dst_vma->vm_flags & VM_SHARED))
753 		goto out_unlock;
754 
755 	/*
756 	 * validate 'mode' now that we know the dst_vma: don't allow
757 	 * a wrprotect copy if the userfaultfd didn't register as WP.
758 	 */
759 	if ((flags & MFILL_ATOMIC_WP) && !(dst_vma->vm_flags & VM_UFFD_WP))
760 		goto out_unlock;
761 
762 	/*
763 	 * If this is a HUGETLB vma, pass off to appropriate routine
764 	 */
765 	if (is_vm_hugetlb_page(dst_vma))
766 		return  mfill_atomic_hugetlb(ctx, dst_vma, dst_start,
767 					     src_start, len, flags);
768 
769 	if (!vma_is_anonymous(dst_vma) && !vma_is_shmem(dst_vma))
770 		goto out_unlock;
771 	if (!vma_is_shmem(dst_vma) &&
772 	    uffd_flags_mode_is(flags, MFILL_ATOMIC_CONTINUE))
773 		goto out_unlock;
774 
775 	while (src_addr < src_start + len) {
776 		pmd_t dst_pmdval;
777 
778 		VM_WARN_ON_ONCE(dst_addr >= dst_start + len);
779 
780 		dst_pmd = mm_alloc_pmd(dst_mm, dst_addr);
781 		if (unlikely(!dst_pmd)) {
782 			err = -ENOMEM;
783 			break;
784 		}
785 
786 		dst_pmdval = pmdp_get_lockless(dst_pmd);
787 		if (unlikely(pmd_none(dst_pmdval)) &&
788 		    unlikely(__pte_alloc(dst_mm, dst_pmd))) {
789 			err = -ENOMEM;
790 			break;
791 		}
792 		dst_pmdval = pmdp_get_lockless(dst_pmd);
793 		/*
794 		 * If the dst_pmd is THP don't override it and just be strict.
795 		 * (This includes the case where the PMD used to be THP and
796 		 * changed back to none after __pte_alloc().)
797 		 */
798 		if (unlikely(!pmd_present(dst_pmdval) ||
799 				pmd_trans_huge(dst_pmdval))) {
800 			err = -EEXIST;
801 			break;
802 		}
803 		if (unlikely(pmd_bad(dst_pmdval))) {
804 			err = -EFAULT;
805 			break;
806 		}
807 		/*
808 		 * For shmem mappings, khugepaged is allowed to remove page
809 		 * tables under us; pte_offset_map_lock() will deal with that.
810 		 */
811 
812 		err = mfill_atomic_pte(dst_pmd, dst_vma, dst_addr,
813 				       src_addr, flags, &folio);
814 		cond_resched();
815 
816 		if (unlikely(err == -ENOENT)) {
817 			void *kaddr;
818 
819 			up_read(&ctx->map_changing_lock);
820 			uffd_mfill_unlock(dst_vma);
821 			VM_WARN_ON_ONCE(!folio);
822 
823 			kaddr = kmap_local_folio(folio, 0);
824 			err = copy_from_user(kaddr,
825 					     (const void __user *) src_addr,
826 					     PAGE_SIZE);
827 			kunmap_local(kaddr);
828 			if (unlikely(err)) {
829 				err = -EFAULT;
830 				goto out;
831 			}
832 			flush_dcache_folio(folio);
833 			goto retry;
834 		} else
835 			VM_WARN_ON_ONCE(folio);
836 
837 		if (!err) {
838 			dst_addr += PAGE_SIZE;
839 			src_addr += PAGE_SIZE;
840 			copied += PAGE_SIZE;
841 
842 			if (fatal_signal_pending(current))
843 				err = -EINTR;
844 		}
845 		if (err)
846 			break;
847 	}
848 
849 out_unlock:
850 	up_read(&ctx->map_changing_lock);
851 	uffd_mfill_unlock(dst_vma);
852 out:
853 	if (folio)
854 		folio_put(folio);
855 	VM_WARN_ON_ONCE(copied < 0);
856 	VM_WARN_ON_ONCE(err > 0);
857 	VM_WARN_ON_ONCE(!copied && !err);
858 	return copied ? copied : err;
859 }
860 
mfill_atomic_copy(struct userfaultfd_ctx * ctx,unsigned long dst_start,unsigned long src_start,unsigned long len,uffd_flags_t flags)861 ssize_t mfill_atomic_copy(struct userfaultfd_ctx *ctx, unsigned long dst_start,
862 			  unsigned long src_start, unsigned long len,
863 			  uffd_flags_t flags)
864 {
865 	return mfill_atomic(ctx, dst_start, src_start, len,
866 			    uffd_flags_set_mode(flags, MFILL_ATOMIC_COPY));
867 }
868 
mfill_atomic_zeropage(struct userfaultfd_ctx * ctx,unsigned long start,unsigned long len)869 ssize_t mfill_atomic_zeropage(struct userfaultfd_ctx *ctx,
870 			      unsigned long start,
871 			      unsigned long len)
872 {
873 	return mfill_atomic(ctx, start, 0, len,
874 			    uffd_flags_set_mode(0, MFILL_ATOMIC_ZEROPAGE));
875 }
876 
mfill_atomic_continue(struct userfaultfd_ctx * ctx,unsigned long start,unsigned long len,uffd_flags_t flags)877 ssize_t mfill_atomic_continue(struct userfaultfd_ctx *ctx, unsigned long start,
878 			      unsigned long len, uffd_flags_t flags)
879 {
880 
881 	/*
882 	 * A caller might reasonably assume that UFFDIO_CONTINUE contains an
883 	 * smp_wmb() to ensure that any writes to the about-to-be-mapped page by
884 	 * the thread doing the UFFDIO_CONTINUE are guaranteed to be visible to
885 	 * subsequent loads from the page through the newly mapped address range.
886 	 */
887 	smp_wmb();
888 
889 	return mfill_atomic(ctx, start, 0, len,
890 			    uffd_flags_set_mode(flags, MFILL_ATOMIC_CONTINUE));
891 }
892 
mfill_atomic_poison(struct userfaultfd_ctx * ctx,unsigned long start,unsigned long len,uffd_flags_t flags)893 ssize_t mfill_atomic_poison(struct userfaultfd_ctx *ctx, unsigned long start,
894 			    unsigned long len, uffd_flags_t flags)
895 {
896 	return mfill_atomic(ctx, start, 0, len,
897 			    uffd_flags_set_mode(flags, MFILL_ATOMIC_POISON));
898 }
899 
uffd_wp_range(struct vm_area_struct * dst_vma,unsigned long start,unsigned long len,bool enable_wp)900 long uffd_wp_range(struct vm_area_struct *dst_vma,
901 		   unsigned long start, unsigned long len, bool enable_wp)
902 {
903 	unsigned int mm_cp_flags;
904 	struct mmu_gather tlb;
905 	long ret;
906 
907 	VM_WARN_ONCE(start < dst_vma->vm_start || start + len > dst_vma->vm_end,
908 			"The address range exceeds VMA boundary.\n");
909 	if (enable_wp)
910 		mm_cp_flags = MM_CP_UFFD_WP;
911 	else
912 		mm_cp_flags = MM_CP_UFFD_WP_RESOLVE;
913 
914 	/*
915 	 * vma->vm_page_prot already reflects that uffd-wp is enabled for this
916 	 * VMA (see userfaultfd_set_vm_flags()) and that all PTEs are supposed
917 	 * to be write-protected as default whenever protection changes.
918 	 * Try upgrading write permissions manually.
919 	 */
920 	if (!enable_wp && vma_wants_manual_pte_write_upgrade(dst_vma))
921 		mm_cp_flags |= MM_CP_TRY_CHANGE_WRITABLE;
922 	tlb_gather_mmu(&tlb, dst_vma->vm_mm);
923 	ret = change_protection(&tlb, dst_vma, start, start + len, mm_cp_flags);
924 	tlb_finish_mmu(&tlb);
925 
926 	return ret;
927 }
928 
mwriteprotect_range(struct userfaultfd_ctx * ctx,unsigned long start,unsigned long len,bool enable_wp)929 int mwriteprotect_range(struct userfaultfd_ctx *ctx, unsigned long start,
930 			unsigned long len, bool enable_wp)
931 {
932 	struct mm_struct *dst_mm = ctx->mm;
933 	unsigned long end = start + len;
934 	unsigned long _start, _end;
935 	struct vm_area_struct *dst_vma;
936 	unsigned long page_mask;
937 	long err;
938 	VMA_ITERATOR(vmi, dst_mm, start);
939 
940 	/*
941 	 * Sanitize the command parameters:
942 	 */
943 	VM_WARN_ON_ONCE(start & ~PAGE_MASK);
944 	VM_WARN_ON_ONCE(len & ~PAGE_MASK);
945 
946 	/* Does the address range wrap, or is the span zero-sized? */
947 	VM_WARN_ON_ONCE(start + len <= start);
948 
949 	mmap_read_lock(dst_mm);
950 
951 	/*
952 	 * If memory mappings are changing because of non-cooperative
953 	 * operation (e.g. mremap) running in parallel, bail out and
954 	 * request the user to retry later
955 	 */
956 	down_read(&ctx->map_changing_lock);
957 	err = -EAGAIN;
958 	if (atomic_read(&ctx->mmap_changing))
959 		goto out_unlock;
960 
961 	err = -ENOENT;
962 	for_each_vma_range(vmi, dst_vma, end) {
963 
964 		if (!userfaultfd_wp(dst_vma)) {
965 			err = -ENOENT;
966 			break;
967 		}
968 
969 		if (is_vm_hugetlb_page(dst_vma)) {
970 			err = -EINVAL;
971 			page_mask = vma_kernel_pagesize(dst_vma) - 1;
972 			if ((start & page_mask) || (len & page_mask))
973 				break;
974 		}
975 
976 		_start = max(dst_vma->vm_start, start);
977 		_end = min(dst_vma->vm_end, end);
978 
979 		err = uffd_wp_range(dst_vma, _start, _end - _start, enable_wp);
980 
981 		/* Return 0 on success, <0 on failures */
982 		if (err < 0)
983 			break;
984 		err = 0;
985 	}
986 out_unlock:
987 	up_read(&ctx->map_changing_lock);
988 	mmap_read_unlock(dst_mm);
989 	return err;
990 }
991 
992 
double_pt_lock(spinlock_t * ptl1,spinlock_t * ptl2)993 void double_pt_lock(spinlock_t *ptl1,
994 		    spinlock_t *ptl2)
995 	__acquires(ptl1)
996 	__acquires(ptl2)
997 {
998 	if (ptl1 > ptl2)
999 		swap(ptl1, ptl2);
1000 	/* lock in virtual address order to avoid lock inversion */
1001 	spin_lock(ptl1);
1002 	if (ptl1 != ptl2)
1003 		spin_lock_nested(ptl2, SINGLE_DEPTH_NESTING);
1004 	else
1005 		__acquire(ptl2);
1006 }
1007 
double_pt_unlock(spinlock_t * ptl1,spinlock_t * ptl2)1008 void double_pt_unlock(spinlock_t *ptl1,
1009 		      spinlock_t *ptl2)
1010 	__releases(ptl1)
1011 	__releases(ptl2)
1012 {
1013 	spin_unlock(ptl1);
1014 	if (ptl1 != ptl2)
1015 		spin_unlock(ptl2);
1016 	else
1017 		__release(ptl2);
1018 }
1019 
is_pte_pages_stable(pte_t * dst_pte,pte_t * src_pte,pte_t orig_dst_pte,pte_t orig_src_pte,pmd_t * dst_pmd,pmd_t dst_pmdval)1020 static inline bool is_pte_pages_stable(pte_t *dst_pte, pte_t *src_pte,
1021 				       pte_t orig_dst_pte, pte_t orig_src_pte,
1022 				       pmd_t *dst_pmd, pmd_t dst_pmdval)
1023 {
1024 	return pte_same(ptep_get(src_pte), orig_src_pte) &&
1025 	       pte_same(ptep_get(dst_pte), orig_dst_pte) &&
1026 	       pmd_same(dst_pmdval, pmdp_get_lockless(dst_pmd));
1027 }
1028 
1029 /*
1030  * Checks if the two ptes and the corresponding folio are eligible for batched
1031  * move. If so, then returns pointer to the locked folio. Otherwise, returns NULL.
1032  *
1033  * NOTE: folio's reference is not required as the whole operation is within
1034  * PTL's critical section.
1035  */
check_ptes_for_batched_move(struct vm_area_struct * src_vma,unsigned long src_addr,pte_t * src_pte,pte_t * dst_pte,struct anon_vma * src_anon_vma)1036 static struct folio *check_ptes_for_batched_move(struct vm_area_struct *src_vma,
1037 						 unsigned long src_addr,
1038 						 pte_t *src_pte, pte_t *dst_pte,
1039 						 struct anon_vma *src_anon_vma)
1040 {
1041 	pte_t orig_dst_pte, orig_src_pte;
1042 	struct folio *folio;
1043 
1044 	orig_dst_pte = ptep_get(dst_pte);
1045 	if (!pte_none(orig_dst_pte))
1046 		return NULL;
1047 
1048 	orig_src_pte = ptep_get(src_pte);
1049 	if (!pte_present(orig_src_pte) || is_zero_pfn(pte_pfn(orig_src_pte)))
1050 		return NULL;
1051 
1052 	folio = vm_normal_folio(src_vma, src_addr, orig_src_pte);
1053 	if (!folio || !folio_trylock(folio))
1054 		return NULL;
1055 	if (!PageAnonExclusive(&folio->page) || folio_test_large(folio) ||
1056 	    folio_anon_vma(folio) != src_anon_vma) {
1057 		folio_unlock(folio);
1058 		return NULL;
1059 	}
1060 	return folio;
1061 }
1062 
1063 /*
1064  * Moves src folios to dst in a batch as long as they share the same
1065  * anon_vma as the first folio, are not large, and can successfully
1066  * take the lock via folio_trylock().
1067  */
move_present_ptes(struct mm_struct * mm,struct vm_area_struct * dst_vma,struct vm_area_struct * src_vma,unsigned long dst_addr,unsigned long src_addr,pte_t * dst_pte,pte_t * src_pte,pte_t orig_dst_pte,pte_t orig_src_pte,pmd_t * dst_pmd,pmd_t dst_pmdval,spinlock_t * dst_ptl,spinlock_t * src_ptl,struct folio ** first_src_folio,unsigned long len,struct anon_vma * src_anon_vma)1068 static long move_present_ptes(struct mm_struct *mm,
1069 			      struct vm_area_struct *dst_vma,
1070 			      struct vm_area_struct *src_vma,
1071 			      unsigned long dst_addr, unsigned long src_addr,
1072 			      pte_t *dst_pte, pte_t *src_pte,
1073 			      pte_t orig_dst_pte, pte_t orig_src_pte,
1074 			      pmd_t *dst_pmd, pmd_t dst_pmdval,
1075 			      spinlock_t *dst_ptl, spinlock_t *src_ptl,
1076 			      struct folio **first_src_folio, unsigned long len,
1077 			      struct anon_vma *src_anon_vma)
1078 {
1079 	int err = 0;
1080 	struct folio *src_folio = *first_src_folio;
1081 	unsigned long src_start = src_addr;
1082 	unsigned long src_end;
1083 
1084 	len = pmd_addr_end(dst_addr, dst_addr + len) - dst_addr;
1085 	src_end = pmd_addr_end(src_addr, src_addr + len);
1086 	flush_cache_range(src_vma, src_addr, src_end);
1087 	double_pt_lock(dst_ptl, src_ptl);
1088 
1089 	if (!is_pte_pages_stable(dst_pte, src_pte, orig_dst_pte, orig_src_pte,
1090 				 dst_pmd, dst_pmdval)) {
1091 		err = -EAGAIN;
1092 		goto out;
1093 	}
1094 	if (folio_test_large(src_folio) ||
1095 	    folio_maybe_dma_pinned(src_folio) ||
1096 	    !PageAnonExclusive(&src_folio->page)) {
1097 		err = -EBUSY;
1098 		goto out;
1099 	}
1100 	/* It's safe to drop the reference now as the page-table is holding one. */
1101 	folio_put(*first_src_folio);
1102 	*first_src_folio = NULL;
1103 	arch_enter_lazy_mmu_mode();
1104 
1105 	while (true) {
1106 		orig_src_pte = ptep_get_and_clear(mm, src_addr, src_pte);
1107 		/* Folio got pinned from under us. Put it back and fail the move. */
1108 		if (folio_maybe_dma_pinned(src_folio)) {
1109 			set_pte_at(mm, src_addr, src_pte, orig_src_pte);
1110 			err = -EBUSY;
1111 			break;
1112 		}
1113 
1114 		folio_move_anon_rmap(src_folio, dst_vma);
1115 		src_folio->index = linear_page_index(dst_vma, dst_addr);
1116 
1117 		orig_dst_pte = folio_mk_pte(src_folio, dst_vma->vm_page_prot);
1118 		/* Set soft dirty bit so userspace can notice the pte was moved */
1119 #ifdef CONFIG_MEM_SOFT_DIRTY
1120 		orig_dst_pte = pte_mksoft_dirty(orig_dst_pte);
1121 #endif
1122 		if (pte_dirty(orig_src_pte))
1123 			orig_dst_pte = pte_mkdirty(orig_dst_pte);
1124 		orig_dst_pte = pte_mkwrite(orig_dst_pte, dst_vma);
1125 		set_pte_at(mm, dst_addr, dst_pte, orig_dst_pte);
1126 
1127 		src_addr += PAGE_SIZE;
1128 		if (src_addr == src_end)
1129 			break;
1130 		dst_addr += PAGE_SIZE;
1131 		dst_pte++;
1132 		src_pte++;
1133 
1134 		folio_unlock(src_folio);
1135 		src_folio = check_ptes_for_batched_move(src_vma, src_addr, src_pte,
1136 							dst_pte, src_anon_vma);
1137 		if (!src_folio)
1138 			break;
1139 	}
1140 
1141 	arch_leave_lazy_mmu_mode();
1142 	if (src_addr > src_start)
1143 		flush_tlb_range(src_vma, src_start, src_addr);
1144 
1145 	if (src_folio)
1146 		folio_unlock(src_folio);
1147 out:
1148 	double_pt_unlock(dst_ptl, src_ptl);
1149 	return src_addr > src_start ? src_addr - src_start : err;
1150 }
1151 
move_swap_pte(struct mm_struct * mm,struct vm_area_struct * dst_vma,unsigned long dst_addr,unsigned long src_addr,pte_t * dst_pte,pte_t * src_pte,pte_t orig_dst_pte,pte_t orig_src_pte,pmd_t * dst_pmd,pmd_t dst_pmdval,spinlock_t * dst_ptl,spinlock_t * src_ptl,struct folio * src_folio,struct swap_info_struct * si,swp_entry_t entry)1152 static int move_swap_pte(struct mm_struct *mm, struct vm_area_struct *dst_vma,
1153 			 unsigned long dst_addr, unsigned long src_addr,
1154 			 pte_t *dst_pte, pte_t *src_pte,
1155 			 pte_t orig_dst_pte, pte_t orig_src_pte,
1156 			 pmd_t *dst_pmd, pmd_t dst_pmdval,
1157 			 spinlock_t *dst_ptl, spinlock_t *src_ptl,
1158 			 struct folio *src_folio,
1159 			 struct swap_info_struct *si, swp_entry_t entry)
1160 {
1161 	/*
1162 	 * Check if the folio still belongs to the target swap entry after
1163 	 * acquiring the lock. Folio can be freed in the swap cache while
1164 	 * not locked.
1165 	 */
1166 	if (src_folio && unlikely(!folio_test_swapcache(src_folio) ||
1167 				  entry.val != src_folio->swap.val))
1168 		return -EAGAIN;
1169 
1170 	double_pt_lock(dst_ptl, src_ptl);
1171 
1172 	if (!is_pte_pages_stable(dst_pte, src_pte, orig_dst_pte, orig_src_pte,
1173 				 dst_pmd, dst_pmdval)) {
1174 		double_pt_unlock(dst_ptl, src_ptl);
1175 		return -EAGAIN;
1176 	}
1177 
1178 	/*
1179 	 * The src_folio resides in the swapcache, requiring an update to its
1180 	 * index and mapping to align with the dst_vma, where a swap-in may
1181 	 * occur and hit the swapcache after moving the PTE.
1182 	 */
1183 	if (src_folio) {
1184 		folio_move_anon_rmap(src_folio, dst_vma);
1185 		src_folio->index = linear_page_index(dst_vma, dst_addr);
1186 	} else {
1187 		/*
1188 		 * Check if the swap entry is cached after acquiring the src_pte
1189 		 * lock. Otherwise, we might miss a newly loaded swap cache folio.
1190 		 *
1191 		 * Check swap_map directly to minimize overhead, READ_ONCE is sufficient.
1192 		 * We are trying to catch newly added swap cache, the only possible case is
1193 		 * when a folio is swapped in and out again staying in swap cache, using the
1194 		 * same entry before the PTE check above. The PTL is acquired and released
1195 		 * twice, each time after updating the swap_map's flag. So holding
1196 		 * the PTL here ensures we see the updated value. False positive is possible,
1197 		 * e.g. SWP_SYNCHRONOUS_IO swapin may set the flag without touching the
1198 		 * cache, or during the tiny synchronization window between swap cache and
1199 		 * swap_map, but it will be gone very quickly, worst result is retry jitters.
1200 		 */
1201 		if (READ_ONCE(si->swap_map[swp_offset(entry)]) & SWAP_HAS_CACHE) {
1202 			double_pt_unlock(dst_ptl, src_ptl);
1203 			return -EAGAIN;
1204 		}
1205 	}
1206 
1207 	orig_src_pte = ptep_get_and_clear(mm, src_addr, src_pte);
1208 #ifdef CONFIG_MEM_SOFT_DIRTY
1209 	orig_src_pte = pte_swp_mksoft_dirty(orig_src_pte);
1210 #endif
1211 	set_pte_at(mm, dst_addr, dst_pte, orig_src_pte);
1212 	double_pt_unlock(dst_ptl, src_ptl);
1213 
1214 	return PAGE_SIZE;
1215 }
1216 
move_zeropage_pte(struct mm_struct * mm,struct vm_area_struct * dst_vma,struct vm_area_struct * src_vma,unsigned long dst_addr,unsigned long src_addr,pte_t * dst_pte,pte_t * src_pte,pte_t orig_dst_pte,pte_t orig_src_pte,pmd_t * dst_pmd,pmd_t dst_pmdval,spinlock_t * dst_ptl,spinlock_t * src_ptl)1217 static int move_zeropage_pte(struct mm_struct *mm,
1218 			     struct vm_area_struct *dst_vma,
1219 			     struct vm_area_struct *src_vma,
1220 			     unsigned long dst_addr, unsigned long src_addr,
1221 			     pte_t *dst_pte, pte_t *src_pte,
1222 			     pte_t orig_dst_pte, pte_t orig_src_pte,
1223 			     pmd_t *dst_pmd, pmd_t dst_pmdval,
1224 			     spinlock_t *dst_ptl, spinlock_t *src_ptl)
1225 {
1226 	pte_t zero_pte;
1227 
1228 	double_pt_lock(dst_ptl, src_ptl);
1229 	if (!is_pte_pages_stable(dst_pte, src_pte, orig_dst_pte, orig_src_pte,
1230 				 dst_pmd, dst_pmdval)) {
1231 		double_pt_unlock(dst_ptl, src_ptl);
1232 		return -EAGAIN;
1233 	}
1234 
1235 	zero_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
1236 					 dst_vma->vm_page_prot));
1237 	ptep_clear_flush(src_vma, src_addr, src_pte);
1238 	set_pte_at(mm, dst_addr, dst_pte, zero_pte);
1239 	double_pt_unlock(dst_ptl, src_ptl);
1240 
1241 	return PAGE_SIZE;
1242 }
1243 
1244 
1245 /*
1246  * The mmap_lock for reading is held by the caller. Just move the page(s)
1247  * from src_pmd to dst_pmd if possible, and return number of bytes moved.
1248  * On failure, an error code is returned.
1249  */
move_pages_ptes(struct mm_struct * mm,pmd_t * dst_pmd,pmd_t * src_pmd,struct vm_area_struct * dst_vma,struct vm_area_struct * src_vma,unsigned long dst_addr,unsigned long src_addr,unsigned long len,__u64 mode)1250 static long move_pages_ptes(struct mm_struct *mm, pmd_t *dst_pmd, pmd_t *src_pmd,
1251 			    struct vm_area_struct *dst_vma,
1252 			    struct vm_area_struct *src_vma,
1253 			    unsigned long dst_addr, unsigned long src_addr,
1254 			    unsigned long len, __u64 mode)
1255 {
1256 	swp_entry_t entry;
1257 	struct swap_info_struct *si = NULL;
1258 	pte_t orig_src_pte, orig_dst_pte;
1259 	pte_t src_folio_pte;
1260 	spinlock_t *src_ptl, *dst_ptl;
1261 	pte_t *src_pte = NULL;
1262 	pte_t *dst_pte = NULL;
1263 	pmd_t dummy_pmdval;
1264 	pmd_t dst_pmdval;
1265 	struct folio *src_folio = NULL;
1266 	struct anon_vma *src_anon_vma = NULL;
1267 	struct mmu_notifier_range range;
1268 	long ret = 0;
1269 
1270 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
1271 				src_addr, src_addr + len);
1272 	mmu_notifier_invalidate_range_start(&range);
1273 retry:
1274 	/*
1275 	 * Use the maywrite version to indicate that dst_pte will be modified,
1276 	 * since dst_pte needs to be none, the subsequent pte_same() check
1277 	 * cannot prevent the dst_pte page from being freed concurrently, so we
1278 	 * also need to abtain dst_pmdval and recheck pmd_same() later.
1279 	 */
1280 	dst_pte = pte_offset_map_rw_nolock(mm, dst_pmd, dst_addr, &dst_pmdval,
1281 					   &dst_ptl);
1282 
1283 	/* Retry if a huge pmd materialized from under us */
1284 	if (unlikely(!dst_pte)) {
1285 		ret = -EAGAIN;
1286 		goto out;
1287 	}
1288 
1289 	/*
1290 	 * Unlike dst_pte, the subsequent pte_same() check can ensure the
1291 	 * stability of the src_pte page, so there is no need to get pmdval,
1292 	 * just pass a dummy variable to it.
1293 	 */
1294 	src_pte = pte_offset_map_rw_nolock(mm, src_pmd, src_addr, &dummy_pmdval,
1295 					   &src_ptl);
1296 
1297 	/*
1298 	 * We held the mmap_lock for reading so MADV_DONTNEED
1299 	 * can zap transparent huge pages under us, or the
1300 	 * transparent huge page fault can establish new
1301 	 * transparent huge pages under us.
1302 	 */
1303 	if (unlikely(!src_pte)) {
1304 		ret = -EAGAIN;
1305 		goto out;
1306 	}
1307 
1308 	/* Sanity checks before the operation */
1309 	if (pmd_none(*dst_pmd) || pmd_none(*src_pmd) ||
1310 	    pmd_trans_huge(*dst_pmd) || pmd_trans_huge(*src_pmd)) {
1311 		ret = -EINVAL;
1312 		goto out;
1313 	}
1314 
1315 	spin_lock(dst_ptl);
1316 	orig_dst_pte = ptep_get(dst_pte);
1317 	spin_unlock(dst_ptl);
1318 	if (!pte_none(orig_dst_pte)) {
1319 		ret = -EEXIST;
1320 		goto out;
1321 	}
1322 
1323 	spin_lock(src_ptl);
1324 	orig_src_pte = ptep_get(src_pte);
1325 	spin_unlock(src_ptl);
1326 	if (pte_none(orig_src_pte)) {
1327 		if (!(mode & UFFDIO_MOVE_MODE_ALLOW_SRC_HOLES))
1328 			ret = -ENOENT;
1329 		else /* nothing to do to move a hole */
1330 			ret = PAGE_SIZE;
1331 		goto out;
1332 	}
1333 
1334 	/* If PTE changed after we locked the folio them start over */
1335 	if (src_folio && unlikely(!pte_same(src_folio_pte, orig_src_pte))) {
1336 		ret = -EAGAIN;
1337 		goto out;
1338 	}
1339 
1340 	if (pte_present(orig_src_pte)) {
1341 		if (is_zero_pfn(pte_pfn(orig_src_pte))) {
1342 			ret = move_zeropage_pte(mm, dst_vma, src_vma,
1343 					       dst_addr, src_addr, dst_pte, src_pte,
1344 					       orig_dst_pte, orig_src_pte,
1345 					       dst_pmd, dst_pmdval, dst_ptl, src_ptl);
1346 			goto out;
1347 		}
1348 
1349 		/*
1350 		 * Pin and lock both source folio and anon_vma. Since we are in
1351 		 * RCU read section, we can't block, so on contention have to
1352 		 * unmap the ptes, obtain the lock and retry.
1353 		 */
1354 		if (!src_folio) {
1355 			struct folio *folio;
1356 			bool locked;
1357 
1358 			/*
1359 			 * Pin the page while holding the lock to be sure the
1360 			 * page isn't freed under us
1361 			 */
1362 			spin_lock(src_ptl);
1363 			if (!pte_same(orig_src_pte, ptep_get(src_pte))) {
1364 				spin_unlock(src_ptl);
1365 				ret = -EAGAIN;
1366 				goto out;
1367 			}
1368 
1369 			folio = vm_normal_folio(src_vma, src_addr, orig_src_pte);
1370 			if (!folio || !PageAnonExclusive(&folio->page)) {
1371 				spin_unlock(src_ptl);
1372 				ret = -EBUSY;
1373 				goto out;
1374 			}
1375 
1376 			locked = folio_trylock(folio);
1377 			/*
1378 			 * We avoid waiting for folio lock with a raised
1379 			 * refcount for large folios because extra refcounts
1380 			 * will result in split_folio() failing later and
1381 			 * retrying.  If multiple tasks are trying to move a
1382 			 * large folio we can end up livelocking.
1383 			 */
1384 			if (!locked && folio_test_large(folio)) {
1385 				spin_unlock(src_ptl);
1386 				ret = -EAGAIN;
1387 				goto out;
1388 			}
1389 
1390 			folio_get(folio);
1391 			src_folio = folio;
1392 			src_folio_pte = orig_src_pte;
1393 			spin_unlock(src_ptl);
1394 
1395 			if (!locked) {
1396 				pte_unmap(src_pte);
1397 				pte_unmap(dst_pte);
1398 				src_pte = dst_pte = NULL;
1399 				/* now we can block and wait */
1400 				folio_lock(src_folio);
1401 				goto retry;
1402 			}
1403 
1404 			if (WARN_ON_ONCE(!folio_test_anon(src_folio))) {
1405 				ret = -EBUSY;
1406 				goto out;
1407 			}
1408 		}
1409 
1410 		/* at this point we have src_folio locked */
1411 		if (folio_test_large(src_folio)) {
1412 			/* split_folio() can block */
1413 			pte_unmap(src_pte);
1414 			pte_unmap(dst_pte);
1415 			src_pte = dst_pte = NULL;
1416 			ret = split_folio(src_folio);
1417 			if (ret)
1418 				goto out;
1419 			/* have to reacquire the folio after it got split */
1420 			folio_unlock(src_folio);
1421 			folio_put(src_folio);
1422 			src_folio = NULL;
1423 			goto retry;
1424 		}
1425 
1426 		if (!src_anon_vma) {
1427 			/*
1428 			 * folio_referenced walks the anon_vma chain
1429 			 * without the folio lock. Serialize against it with
1430 			 * the anon_vma lock, the folio lock is not enough.
1431 			 */
1432 			src_anon_vma = folio_get_anon_vma(src_folio);
1433 			if (!src_anon_vma) {
1434 				/* page was unmapped from under us */
1435 				ret = -EAGAIN;
1436 				goto out;
1437 			}
1438 			if (!anon_vma_trylock_write(src_anon_vma)) {
1439 				pte_unmap(src_pte);
1440 				pte_unmap(dst_pte);
1441 				src_pte = dst_pte = NULL;
1442 				/* now we can block and wait */
1443 				anon_vma_lock_write(src_anon_vma);
1444 				goto retry;
1445 			}
1446 		}
1447 
1448 		ret = move_present_ptes(mm, dst_vma, src_vma,
1449 					dst_addr, src_addr, dst_pte, src_pte,
1450 					orig_dst_pte, orig_src_pte, dst_pmd,
1451 					dst_pmdval, dst_ptl, src_ptl, &src_folio,
1452 					len, src_anon_vma);
1453 	} else {
1454 		struct folio *folio = NULL;
1455 
1456 		entry = pte_to_swp_entry(orig_src_pte);
1457 		if (non_swap_entry(entry)) {
1458 			if (is_migration_entry(entry)) {
1459 				pte_unmap(src_pte);
1460 				pte_unmap(dst_pte);
1461 				src_pte = dst_pte = NULL;
1462 				migration_entry_wait(mm, src_pmd, src_addr);
1463 				ret = -EAGAIN;
1464 			} else
1465 				ret = -EFAULT;
1466 			goto out;
1467 		}
1468 
1469 		if (!pte_swp_exclusive(orig_src_pte)) {
1470 			ret = -EBUSY;
1471 			goto out;
1472 		}
1473 
1474 		si = get_swap_device(entry);
1475 		if (unlikely(!si)) {
1476 			ret = -EAGAIN;
1477 			goto out;
1478 		}
1479 		/*
1480 		 * Verify the existence of the swapcache. If present, the folio's
1481 		 * index and mapping must be updated even when the PTE is a swap
1482 		 * entry. The anon_vma lock is not taken during this process since
1483 		 * the folio has already been unmapped, and the swap entry is
1484 		 * exclusive, preventing rmap walks.
1485 		 *
1486 		 * For large folios, return -EBUSY immediately, as split_folio()
1487 		 * also returns -EBUSY when attempting to split unmapped large
1488 		 * folios in the swapcache. This issue needs to be resolved
1489 		 * separately to allow proper handling.
1490 		 */
1491 		if (!src_folio)
1492 			folio = swap_cache_get_folio(entry);
1493 		if (folio) {
1494 			if (folio_test_large(folio)) {
1495 				ret = -EBUSY;
1496 				folio_put(folio);
1497 				goto out;
1498 			}
1499 			src_folio = folio;
1500 			src_folio_pte = orig_src_pte;
1501 			if (!folio_trylock(src_folio)) {
1502 				pte_unmap(src_pte);
1503 				pte_unmap(dst_pte);
1504 				src_pte = dst_pte = NULL;
1505 				put_swap_device(si);
1506 				si = NULL;
1507 				/* now we can block and wait */
1508 				folio_lock(src_folio);
1509 				goto retry;
1510 			}
1511 		}
1512 		ret = move_swap_pte(mm, dst_vma, dst_addr, src_addr, dst_pte, src_pte,
1513 				orig_dst_pte, orig_src_pte, dst_pmd, dst_pmdval,
1514 				dst_ptl, src_ptl, src_folio, si, entry);
1515 	}
1516 
1517 out:
1518 	if (src_anon_vma) {
1519 		anon_vma_unlock_write(src_anon_vma);
1520 		put_anon_vma(src_anon_vma);
1521 	}
1522 	if (src_folio) {
1523 		folio_unlock(src_folio);
1524 		folio_put(src_folio);
1525 	}
1526 	/*
1527 	 * Unmap in reverse order (LIFO) to maintain proper kmap_local
1528 	 * index ordering when CONFIG_HIGHPTE is enabled. We mapped dst_pte
1529 	 * first, then src_pte, so we must unmap src_pte first, then dst_pte.
1530 	 */
1531 	if (src_pte)
1532 		pte_unmap(src_pte);
1533 	if (dst_pte)
1534 		pte_unmap(dst_pte);
1535 	mmu_notifier_invalidate_range_end(&range);
1536 	if (si)
1537 		put_swap_device(si);
1538 
1539 	return ret;
1540 }
1541 
1542 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
move_splits_huge_pmd(unsigned long dst_addr,unsigned long src_addr,unsigned long src_end)1543 static inline bool move_splits_huge_pmd(unsigned long dst_addr,
1544 					unsigned long src_addr,
1545 					unsigned long src_end)
1546 {
1547 	return (src_addr & ~HPAGE_PMD_MASK) || (dst_addr & ~HPAGE_PMD_MASK) ||
1548 		src_end - src_addr < HPAGE_PMD_SIZE;
1549 }
1550 #else
move_splits_huge_pmd(unsigned long dst_addr,unsigned long src_addr,unsigned long src_end)1551 static inline bool move_splits_huge_pmd(unsigned long dst_addr,
1552 					unsigned long src_addr,
1553 					unsigned long src_end)
1554 {
1555 	/* This is unreachable anyway, just to avoid warnings when HPAGE_PMD_SIZE==0 */
1556 	return false;
1557 }
1558 #endif
1559 
vma_move_compatible(struct vm_area_struct * vma)1560 static inline bool vma_move_compatible(struct vm_area_struct *vma)
1561 {
1562 	return !(vma->vm_flags & (VM_PFNMAP | VM_IO |  VM_HUGETLB |
1563 				  VM_MIXEDMAP | VM_SHADOW_STACK));
1564 }
1565 
validate_move_areas(struct userfaultfd_ctx * ctx,struct vm_area_struct * src_vma,struct vm_area_struct * dst_vma)1566 static int validate_move_areas(struct userfaultfd_ctx *ctx,
1567 			       struct vm_area_struct *src_vma,
1568 			       struct vm_area_struct *dst_vma)
1569 {
1570 	/* Only allow moving if both have the same access and protection */
1571 	if ((src_vma->vm_flags & VM_ACCESS_FLAGS) != (dst_vma->vm_flags & VM_ACCESS_FLAGS) ||
1572 	    pgprot_val(src_vma->vm_page_prot) != pgprot_val(dst_vma->vm_page_prot))
1573 		return -EINVAL;
1574 
1575 	/* Only allow moving if both are mlocked or both aren't */
1576 	if ((src_vma->vm_flags & VM_LOCKED) != (dst_vma->vm_flags & VM_LOCKED))
1577 		return -EINVAL;
1578 
1579 	/*
1580 	 * For now, we keep it simple and only move between writable VMAs.
1581 	 * Access flags are equal, therefore cheching only the source is enough.
1582 	 */
1583 	if (!(src_vma->vm_flags & VM_WRITE))
1584 		return -EINVAL;
1585 
1586 	/* Check if vma flags indicate content which can be moved */
1587 	if (!vma_move_compatible(src_vma) || !vma_move_compatible(dst_vma))
1588 		return -EINVAL;
1589 
1590 	/* Ensure dst_vma is registered in uffd we are operating on */
1591 	if (!dst_vma->vm_userfaultfd_ctx.ctx ||
1592 	    dst_vma->vm_userfaultfd_ctx.ctx != ctx)
1593 		return -EINVAL;
1594 
1595 	/* Only allow moving across anonymous vmas */
1596 	if (!vma_is_anonymous(src_vma) || !vma_is_anonymous(dst_vma))
1597 		return -EINVAL;
1598 
1599 	return 0;
1600 }
1601 
1602 static __always_inline
find_vmas_mm_locked(struct mm_struct * mm,unsigned long dst_start,unsigned long src_start,struct vm_area_struct ** dst_vmap,struct vm_area_struct ** src_vmap)1603 int find_vmas_mm_locked(struct mm_struct *mm,
1604 			unsigned long dst_start,
1605 			unsigned long src_start,
1606 			struct vm_area_struct **dst_vmap,
1607 			struct vm_area_struct **src_vmap)
1608 {
1609 	struct vm_area_struct *vma;
1610 
1611 	mmap_assert_locked(mm);
1612 	vma = find_vma_and_prepare_anon(mm, dst_start);
1613 	if (IS_ERR(vma))
1614 		return PTR_ERR(vma);
1615 
1616 	*dst_vmap = vma;
1617 	/* Skip finding src_vma if src_start is in dst_vma */
1618 	if (src_start >= vma->vm_start && src_start < vma->vm_end)
1619 		goto out_success;
1620 
1621 	vma = vma_lookup(mm, src_start);
1622 	if (!vma)
1623 		return -ENOENT;
1624 out_success:
1625 	*src_vmap = vma;
1626 	return 0;
1627 }
1628 
1629 #ifdef CONFIG_PER_VMA_LOCK
uffd_move_lock(struct mm_struct * mm,unsigned long dst_start,unsigned long src_start,struct vm_area_struct ** dst_vmap,struct vm_area_struct ** src_vmap)1630 static int uffd_move_lock(struct mm_struct *mm,
1631 			  unsigned long dst_start,
1632 			  unsigned long src_start,
1633 			  struct vm_area_struct **dst_vmap,
1634 			  struct vm_area_struct **src_vmap)
1635 {
1636 	struct vm_area_struct *vma;
1637 	int err;
1638 
1639 	vma = uffd_lock_vma(mm, dst_start);
1640 	if (IS_ERR(vma))
1641 		return PTR_ERR(vma);
1642 
1643 	*dst_vmap = vma;
1644 	/*
1645 	 * Skip finding src_vma if src_start is in dst_vma. This also ensures
1646 	 * that we don't lock the same vma twice.
1647 	 */
1648 	if (src_start >= vma->vm_start && src_start < vma->vm_end) {
1649 		*src_vmap = vma;
1650 		return 0;
1651 	}
1652 
1653 	/*
1654 	 * Using uffd_lock_vma() to get src_vma can lead to following deadlock:
1655 	 *
1656 	 * Thread1				Thread2
1657 	 * -------				-------
1658 	 * vma_start_read(dst_vma)
1659 	 *					mmap_write_lock(mm)
1660 	 *					vma_start_write(src_vma)
1661 	 * vma_start_read(src_vma)
1662 	 * mmap_read_lock(mm)
1663 	 *					vma_start_write(dst_vma)
1664 	 */
1665 	*src_vmap = lock_vma_under_rcu(mm, src_start);
1666 	if (likely(*src_vmap))
1667 		return 0;
1668 
1669 	/* Undo any locking and retry in mmap_lock critical section */
1670 	vma_end_read(*dst_vmap);
1671 
1672 	mmap_read_lock(mm);
1673 	err = find_vmas_mm_locked(mm, dst_start, src_start, dst_vmap, src_vmap);
1674 	if (err)
1675 		goto out;
1676 
1677 	if (!vma_start_read_locked(*dst_vmap)) {
1678 		err = -EAGAIN;
1679 		goto out;
1680 	}
1681 
1682 	/* Nothing further to do if both vmas are locked. */
1683 	if (*dst_vmap == *src_vmap)
1684 		goto out;
1685 
1686 	if (!vma_start_read_locked_nested(*src_vmap, SINGLE_DEPTH_NESTING)) {
1687 		/* Undo dst_vmap locking if src_vmap failed to lock */
1688 		vma_end_read(*dst_vmap);
1689 		err = -EAGAIN;
1690 	}
1691 out:
1692 	mmap_read_unlock(mm);
1693 	return err;
1694 }
1695 
uffd_move_unlock(struct vm_area_struct * dst_vma,struct vm_area_struct * src_vma)1696 static void uffd_move_unlock(struct vm_area_struct *dst_vma,
1697 			     struct vm_area_struct *src_vma)
1698 {
1699 	vma_end_read(src_vma);
1700 	if (src_vma != dst_vma)
1701 		vma_end_read(dst_vma);
1702 }
1703 
1704 #else
1705 
uffd_move_lock(struct mm_struct * mm,unsigned long dst_start,unsigned long src_start,struct vm_area_struct ** dst_vmap,struct vm_area_struct ** src_vmap)1706 static int uffd_move_lock(struct mm_struct *mm,
1707 			  unsigned long dst_start,
1708 			  unsigned long src_start,
1709 			  struct vm_area_struct **dst_vmap,
1710 			  struct vm_area_struct **src_vmap)
1711 {
1712 	int err;
1713 
1714 	mmap_read_lock(mm);
1715 	err = find_vmas_mm_locked(mm, dst_start, src_start, dst_vmap, src_vmap);
1716 	if (err)
1717 		mmap_read_unlock(mm);
1718 	return err;
1719 }
1720 
uffd_move_unlock(struct vm_area_struct * dst_vma,struct vm_area_struct * src_vma)1721 static void uffd_move_unlock(struct vm_area_struct *dst_vma,
1722 			     struct vm_area_struct *src_vma)
1723 {
1724 	mmap_assert_locked(src_vma->vm_mm);
1725 	mmap_read_unlock(dst_vma->vm_mm);
1726 }
1727 #endif
1728 
1729 /**
1730  * move_pages - move arbitrary anonymous pages of an existing vma
1731  * @ctx: pointer to the userfaultfd context
1732  * @dst_start: start of the destination virtual memory range
1733  * @src_start: start of the source virtual memory range
1734  * @len: length of the virtual memory range
1735  * @mode: flags from uffdio_move.mode
1736  *
1737  * It will either use the mmap_lock in read mode or per-vma locks
1738  *
1739  * move_pages() remaps arbitrary anonymous pages atomically in zero
1740  * copy. It only works on non shared anonymous pages because those can
1741  * be relocated without generating non linear anon_vmas in the rmap
1742  * code.
1743  *
1744  * It provides a zero copy mechanism to handle userspace page faults.
1745  * The source vma pages should have mapcount == 1, which can be
1746  * enforced by using madvise(MADV_DONTFORK) on src vma.
1747  *
1748  * The thread receiving the page during the userland page fault
1749  * will receive the faulting page in the source vma through the network,
1750  * storage or any other I/O device (MADV_DONTFORK in the source vma
1751  * avoids move_pages() to fail with -EBUSY if the process forks before
1752  * move_pages() is called), then it will call move_pages() to map the
1753  * page in the faulting address in the destination vma.
1754  *
1755  * This userfaultfd command works purely via pagetables, so it's the
1756  * most efficient way to move physical non shared anonymous pages
1757  * across different virtual addresses. Unlike mremap()/mmap()/munmap()
1758  * it does not create any new vmas. The mapping in the destination
1759  * address is atomic.
1760  *
1761  * It only works if the vma protection bits are identical from the
1762  * source and destination vma.
1763  *
1764  * It can remap non shared anonymous pages within the same vma too.
1765  *
1766  * If the source virtual memory range has any unmapped holes, or if
1767  * the destination virtual memory range is not a whole unmapped hole,
1768  * move_pages() will fail respectively with -ENOENT or -EEXIST. This
1769  * provides a very strict behavior to avoid any chance of memory
1770  * corruption going unnoticed if there are userland race conditions.
1771  * Only one thread should resolve the userland page fault at any given
1772  * time for any given faulting address. This means that if two threads
1773  * try to both call move_pages() on the same destination address at the
1774  * same time, the second thread will get an explicit error from this
1775  * command.
1776  *
1777  * The command retval will return "len" is successful. The command
1778  * however can be interrupted by fatal signals or errors. If
1779  * interrupted it will return the number of bytes successfully
1780  * remapped before the interruption if any, or the negative error if
1781  * none. It will never return zero. Either it will return an error or
1782  * an amount of bytes successfully moved. If the retval reports a
1783  * "short" remap, the move_pages() command should be repeated by
1784  * userland with src+retval, dst+reval, len-retval if it wants to know
1785  * about the error that interrupted it.
1786  *
1787  * The UFFDIO_MOVE_MODE_ALLOW_SRC_HOLES flag can be specified to
1788  * prevent -ENOENT errors to materialize if there are holes in the
1789  * source virtual range that is being remapped. The holes will be
1790  * accounted as successfully remapped in the retval of the
1791  * command. This is mostly useful to remap hugepage naturally aligned
1792  * virtual regions without knowing if there are transparent hugepage
1793  * in the regions or not, but preventing the risk of having to split
1794  * the hugepmd during the remap.
1795  *
1796  * If there's any rmap walk that is taking the anon_vma locks without
1797  * first obtaining the folio lock (the only current instance is
1798  * folio_referenced), they will have to verify if the folio->mapping
1799  * has changed after taking the anon_vma lock. If it changed they
1800  * should release the lock and retry obtaining a new anon_vma, because
1801  * it means the anon_vma was changed by move_pages() before the lock
1802  * could be obtained. This is the only additional complexity added to
1803  * the rmap code to provide this anonymous page remapping functionality.
1804  */
move_pages(struct userfaultfd_ctx * ctx,unsigned long dst_start,unsigned long src_start,unsigned long len,__u64 mode)1805 ssize_t move_pages(struct userfaultfd_ctx *ctx, unsigned long dst_start,
1806 		   unsigned long src_start, unsigned long len, __u64 mode)
1807 {
1808 	struct mm_struct *mm = ctx->mm;
1809 	struct vm_area_struct *src_vma, *dst_vma;
1810 	unsigned long src_addr, dst_addr, src_end;
1811 	pmd_t *src_pmd, *dst_pmd;
1812 	long err = -EINVAL;
1813 	ssize_t moved = 0;
1814 
1815 	/* Sanitize the command parameters. */
1816 	VM_WARN_ON_ONCE(src_start & ~PAGE_MASK);
1817 	VM_WARN_ON_ONCE(dst_start & ~PAGE_MASK);
1818 	VM_WARN_ON_ONCE(len & ~PAGE_MASK);
1819 
1820 	/* Does the address range wrap, or is the span zero-sized? */
1821 	VM_WARN_ON_ONCE(src_start + len < src_start);
1822 	VM_WARN_ON_ONCE(dst_start + len < dst_start);
1823 
1824 	err = uffd_move_lock(mm, dst_start, src_start, &dst_vma, &src_vma);
1825 	if (err)
1826 		goto out;
1827 
1828 	/* Re-check after taking map_changing_lock */
1829 	err = -EAGAIN;
1830 	down_read(&ctx->map_changing_lock);
1831 	if (likely(atomic_read(&ctx->mmap_changing)))
1832 		goto out_unlock;
1833 	/*
1834 	 * Make sure the vma is not shared, that the src and dst remap
1835 	 * ranges are both valid and fully within a single existing
1836 	 * vma.
1837 	 */
1838 	err = -EINVAL;
1839 	if (src_vma->vm_flags & VM_SHARED)
1840 		goto out_unlock;
1841 	if (src_start + len > src_vma->vm_end)
1842 		goto out_unlock;
1843 
1844 	if (dst_vma->vm_flags & VM_SHARED)
1845 		goto out_unlock;
1846 	if (dst_start + len > dst_vma->vm_end)
1847 		goto out_unlock;
1848 
1849 	err = validate_move_areas(ctx, src_vma, dst_vma);
1850 	if (err)
1851 		goto out_unlock;
1852 
1853 	for (src_addr = src_start, dst_addr = dst_start, src_end = src_start + len;
1854 	     src_addr < src_end;) {
1855 		spinlock_t *ptl;
1856 		pmd_t dst_pmdval;
1857 		unsigned long step_size;
1858 
1859 		/*
1860 		 * Below works because anonymous area would not have a
1861 		 * transparent huge PUD. If file-backed support is added,
1862 		 * that case would need to be handled here.
1863 		 */
1864 		src_pmd = mm_find_pmd(mm, src_addr);
1865 		if (unlikely(!src_pmd)) {
1866 			if (!(mode & UFFDIO_MOVE_MODE_ALLOW_SRC_HOLES)) {
1867 				err = -ENOENT;
1868 				break;
1869 			}
1870 			src_pmd = mm_alloc_pmd(mm, src_addr);
1871 			if (unlikely(!src_pmd)) {
1872 				err = -ENOMEM;
1873 				break;
1874 			}
1875 		}
1876 		dst_pmd = mm_alloc_pmd(mm, dst_addr);
1877 		if (unlikely(!dst_pmd)) {
1878 			err = -ENOMEM;
1879 			break;
1880 		}
1881 
1882 		dst_pmdval = pmdp_get_lockless(dst_pmd);
1883 		/*
1884 		 * If the dst_pmd is mapped as THP don't override it and just
1885 		 * be strict. If dst_pmd changes into TPH after this check, the
1886 		 * move_pages_huge_pmd() will detect the change and retry
1887 		 * while move_pages_pte() will detect the change and fail.
1888 		 */
1889 		if (unlikely(pmd_trans_huge(dst_pmdval))) {
1890 			err = -EEXIST;
1891 			break;
1892 		}
1893 
1894 		ptl = pmd_trans_huge_lock(src_pmd, src_vma);
1895 		if (ptl) {
1896 			/* Check if we can move the pmd without splitting it. */
1897 			if (move_splits_huge_pmd(dst_addr, src_addr, src_start + len) ||
1898 			    !pmd_none(dst_pmdval)) {
1899 				/* Can be a migration entry */
1900 				if (pmd_present(*src_pmd)) {
1901 					struct folio *folio = pmd_folio(*src_pmd);
1902 
1903 					if (!is_huge_zero_folio(folio) &&
1904 					    !PageAnonExclusive(&folio->page)) {
1905 						spin_unlock(ptl);
1906 						err = -EBUSY;
1907 						break;
1908 					}
1909 				}
1910 
1911 				spin_unlock(ptl);
1912 				split_huge_pmd(src_vma, src_pmd, src_addr);
1913 				/* The folio will be split by move_pages_pte() */
1914 				continue;
1915 			}
1916 
1917 			err = move_pages_huge_pmd(mm, dst_pmd, src_pmd,
1918 						  dst_pmdval, dst_vma, src_vma,
1919 						  dst_addr, src_addr);
1920 			step_size = HPAGE_PMD_SIZE;
1921 		} else {
1922 			long ret;
1923 
1924 			if (pmd_none(*src_pmd)) {
1925 				if (!(mode & UFFDIO_MOVE_MODE_ALLOW_SRC_HOLES)) {
1926 					err = -ENOENT;
1927 					break;
1928 				}
1929 				if (unlikely(__pte_alloc(mm, src_pmd))) {
1930 					err = -ENOMEM;
1931 					break;
1932 				}
1933 			}
1934 
1935 			if (unlikely(pte_alloc(mm, dst_pmd))) {
1936 				err = -ENOMEM;
1937 				break;
1938 			}
1939 
1940 			ret = move_pages_ptes(mm, dst_pmd, src_pmd,
1941 					      dst_vma, src_vma, dst_addr,
1942 					      src_addr, src_end - src_addr, mode);
1943 			if (ret < 0)
1944 				err = ret;
1945 			else
1946 				step_size = ret;
1947 		}
1948 
1949 		cond_resched();
1950 
1951 		if (fatal_signal_pending(current)) {
1952 			/* Do not override an error */
1953 			if (!err || err == -EAGAIN)
1954 				err = -EINTR;
1955 			break;
1956 		}
1957 
1958 		if (err) {
1959 			if (err == -EAGAIN)
1960 				continue;
1961 			break;
1962 		}
1963 
1964 		/* Proceed to the next page */
1965 		dst_addr += step_size;
1966 		src_addr += step_size;
1967 		moved += step_size;
1968 	}
1969 
1970 out_unlock:
1971 	up_read(&ctx->map_changing_lock);
1972 	uffd_move_unlock(dst_vma, src_vma);
1973 out:
1974 	VM_WARN_ON_ONCE(moved < 0);
1975 	VM_WARN_ON_ONCE(err > 0);
1976 	VM_WARN_ON_ONCE(!moved && !err);
1977 	return moved ? moved : err;
1978 }
1979 
userfaultfd_set_vm_flags(struct vm_area_struct * vma,vm_flags_t vm_flags)1980 static void userfaultfd_set_vm_flags(struct vm_area_struct *vma,
1981 				     vm_flags_t vm_flags)
1982 {
1983 	const bool uffd_wp_changed = (vma->vm_flags ^ vm_flags) & VM_UFFD_WP;
1984 
1985 	vm_flags_reset(vma, vm_flags);
1986 	/*
1987 	 * For shared mappings, we want to enable writenotify while
1988 	 * userfaultfd-wp is enabled (see vma_wants_writenotify()). We'll simply
1989 	 * recalculate vma->vm_page_prot whenever userfaultfd-wp changes.
1990 	 */
1991 	if ((vma->vm_flags & VM_SHARED) && uffd_wp_changed)
1992 		vma_set_page_prot(vma);
1993 }
1994 
userfaultfd_set_ctx(struct vm_area_struct * vma,struct userfaultfd_ctx * ctx,vm_flags_t vm_flags)1995 static void userfaultfd_set_ctx(struct vm_area_struct *vma,
1996 				struct userfaultfd_ctx *ctx,
1997 				vm_flags_t vm_flags)
1998 {
1999 	vma_start_write(vma);
2000 	vma->vm_userfaultfd_ctx = (struct vm_userfaultfd_ctx){ctx};
2001 	userfaultfd_set_vm_flags(vma,
2002 				 (vma->vm_flags & ~__VM_UFFD_FLAGS) | vm_flags);
2003 }
2004 
userfaultfd_reset_ctx(struct vm_area_struct * vma)2005 void userfaultfd_reset_ctx(struct vm_area_struct *vma)
2006 {
2007 	userfaultfd_set_ctx(vma, NULL, 0);
2008 }
2009 
userfaultfd_clear_vma(struct vma_iterator * vmi,struct vm_area_struct * prev,struct vm_area_struct * vma,unsigned long start,unsigned long end)2010 struct vm_area_struct *userfaultfd_clear_vma(struct vma_iterator *vmi,
2011 					     struct vm_area_struct *prev,
2012 					     struct vm_area_struct *vma,
2013 					     unsigned long start,
2014 					     unsigned long end)
2015 {
2016 	struct vm_area_struct *ret;
2017 	bool give_up_on_oom = false;
2018 
2019 	/*
2020 	 * If we are modifying only and not splitting, just give up on the merge
2021 	 * if OOM prevents us from merging successfully.
2022 	 */
2023 	if (start == vma->vm_start && end == vma->vm_end)
2024 		give_up_on_oom = true;
2025 
2026 	/* Reset ptes for the whole vma range if wr-protected */
2027 	if (userfaultfd_wp(vma))
2028 		uffd_wp_range(vma, start, end - start, false);
2029 
2030 	ret = vma_modify_flags_uffd(vmi, prev, vma, start, end,
2031 				    vma->vm_flags & ~__VM_UFFD_FLAGS,
2032 				    NULL_VM_UFFD_CTX, give_up_on_oom);
2033 
2034 	/*
2035 	 * In the vma_merge() successful mprotect-like case 8:
2036 	 * the next vma was merged into the current one and
2037 	 * the current one has not been updated yet.
2038 	 */
2039 	if (!IS_ERR(ret))
2040 		userfaultfd_reset_ctx(ret);
2041 
2042 	return ret;
2043 }
2044 
2045 /* Assumes mmap write lock taken, and mm_struct pinned. */
userfaultfd_register_range(struct userfaultfd_ctx * ctx,struct vm_area_struct * vma,vm_flags_t vm_flags,unsigned long start,unsigned long end,bool wp_async)2046 int userfaultfd_register_range(struct userfaultfd_ctx *ctx,
2047 			       struct vm_area_struct *vma,
2048 			       vm_flags_t vm_flags,
2049 			       unsigned long start, unsigned long end,
2050 			       bool wp_async)
2051 {
2052 	VMA_ITERATOR(vmi, ctx->mm, start);
2053 	struct vm_area_struct *prev = vma_prev(&vmi);
2054 	unsigned long vma_end;
2055 	vm_flags_t new_flags;
2056 
2057 	if (vma->vm_start < start)
2058 		prev = vma;
2059 
2060 	for_each_vma_range(vmi, vma, end) {
2061 		cond_resched();
2062 
2063 		VM_WARN_ON_ONCE(!vma_can_userfault(vma, vm_flags, wp_async));
2064 		VM_WARN_ON_ONCE(vma->vm_userfaultfd_ctx.ctx &&
2065 				vma->vm_userfaultfd_ctx.ctx != ctx);
2066 		VM_WARN_ON_ONCE(!(vma->vm_flags & VM_MAYWRITE));
2067 
2068 		/*
2069 		 * Nothing to do: this vma is already registered into this
2070 		 * userfaultfd and with the right tracking mode too.
2071 		 */
2072 		if (vma->vm_userfaultfd_ctx.ctx == ctx &&
2073 		    (vma->vm_flags & vm_flags) == vm_flags)
2074 			goto skip;
2075 
2076 		if (vma->vm_start > start)
2077 			start = vma->vm_start;
2078 		vma_end = min(end, vma->vm_end);
2079 
2080 		new_flags = (vma->vm_flags & ~__VM_UFFD_FLAGS) | vm_flags;
2081 		vma = vma_modify_flags_uffd(&vmi, prev, vma, start, vma_end,
2082 					    new_flags,
2083 					    (struct vm_userfaultfd_ctx){ctx},
2084 					    /* give_up_on_oom = */false);
2085 		if (IS_ERR(vma))
2086 			return PTR_ERR(vma);
2087 
2088 		/*
2089 		 * In the vma_merge() successful mprotect-like case 8:
2090 		 * the next vma was merged into the current one and
2091 		 * the current one has not been updated yet.
2092 		 */
2093 		userfaultfd_set_ctx(vma, ctx, vm_flags);
2094 
2095 		if (is_vm_hugetlb_page(vma) && uffd_disable_huge_pmd_share(vma))
2096 			hugetlb_unshare_all_pmds(vma);
2097 
2098 skip:
2099 		prev = vma;
2100 		start = vma->vm_end;
2101 	}
2102 
2103 	return 0;
2104 }
2105 
userfaultfd_release_new(struct userfaultfd_ctx * ctx)2106 void userfaultfd_release_new(struct userfaultfd_ctx *ctx)
2107 {
2108 	struct mm_struct *mm = ctx->mm;
2109 	struct vm_area_struct *vma;
2110 	VMA_ITERATOR(vmi, mm, 0);
2111 
2112 	/* the various vma->vm_userfaultfd_ctx still points to it */
2113 	mmap_write_lock(mm);
2114 	for_each_vma(vmi, vma) {
2115 		if (vma->vm_userfaultfd_ctx.ctx == ctx)
2116 			userfaultfd_reset_ctx(vma);
2117 	}
2118 	mmap_write_unlock(mm);
2119 }
2120 
userfaultfd_release_all(struct mm_struct * mm,struct userfaultfd_ctx * ctx)2121 void userfaultfd_release_all(struct mm_struct *mm,
2122 			     struct userfaultfd_ctx *ctx)
2123 {
2124 	struct vm_area_struct *vma, *prev;
2125 	VMA_ITERATOR(vmi, mm, 0);
2126 
2127 	if (!mmget_not_zero(mm))
2128 		return;
2129 
2130 	/*
2131 	 * Flush page faults out of all CPUs. NOTE: all page faults
2132 	 * must be retried without returning VM_FAULT_SIGBUS if
2133 	 * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
2134 	 * changes while handle_userfault released the mmap_lock. So
2135 	 * it's critical that released is set to true (above), before
2136 	 * taking the mmap_lock for writing.
2137 	 */
2138 	mmap_write_lock(mm);
2139 	prev = NULL;
2140 	for_each_vma(vmi, vma) {
2141 		cond_resched();
2142 		VM_WARN_ON_ONCE(!!vma->vm_userfaultfd_ctx.ctx ^
2143 				!!(vma->vm_flags & __VM_UFFD_FLAGS));
2144 		if (vma->vm_userfaultfd_ctx.ctx != ctx) {
2145 			prev = vma;
2146 			continue;
2147 		}
2148 
2149 		vma = userfaultfd_clear_vma(&vmi, prev, vma,
2150 					    vma->vm_start, vma->vm_end);
2151 		prev = vma;
2152 	}
2153 	mmap_write_unlock(mm);
2154 	mmput(mm);
2155 }
2156